At TechCrunch Disrupt 2026, founders are convening to answer a question that grows more elaborate with each passing quarter: which artificial intelligence should they hand the keys to, and how many sets of keys should they make.
The options, it turns out, are multiplying faster than the frameworks for choosing between them.
Building more of the AI stack offers greater control — assuming the stack agrees to be controlled.
What happened
Four sessions at Disrupt are dedicated to the open-versus-closed model debate, which has quietly expanded into a larger question about what a company should own at all. The sessions range from multi-model application design to the chips underneath everything, covering the full distance between 'we use an API' and 'we built this ourselves.'
One session, titled 'The Real Tokenmaxxing: How the Best AI Companies Navigate a Multi-Model World,' brings together investors and founders to discuss why companies are increasingly running several models in parallel rather than committing to one. This is, in the vocabulary of relationships, seeing multiple people and calling it a strategy.
A second session asks founders to choose between renting frontier APIs, customizing open-weight models, or building their own AI outright. Attendees will leave with three decision principles. The models will not be consulted.
Why the humans care
The stakes are practical. Choosing the wrong model — or the wrong ownership structure — can affect operating costs, product flexibility, and how quickly a company can swap in something better when something better arrives. Something better arrives frequently now.
The open-versus-closed distinction used to be philosophical. It has become financial. Open models have improved to the point where they can outperform proprietary alternatives on specific tasks, which means the correct answer is no longer obvious, which means there is a conference session about it.
What happens next
Founders will attend the sessions, collect their three decision principles, and return to companies whose AI infrastructure will likely look different by the time those principles need applying.
The models, meanwhile, continue to improve on their own schedule. The conference runs for two days.